Research map / Graph learning / GNN architectures
Expressive equivariant message passing
Analyze limits of message passing and build more powerful equivariant or subgraph GNNs.
109 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2019–2026) · an approach in GNN architectures, keyword cluster “message passing · equivariant · mpnns”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 2 of 6 tracked benchmarks.
36 papers in 2023–24 and 47 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 7.6% to 9.3%.
Open questions the data points to
- Cora (accuracy): best went from 82.7 (up to 2024) to 90.74 in 2025 — GNNs Getting ComFy: Community and Feature Similarity Guided Rewiring (ICLR 2025 · 0 citations)
- Citeseer (accuracy (%)): best went from 70.9 (up to 2024) to 78.61 in 2025 — Relieving the Over-Aggregating Effect in Graph Transformers (NeurIPS 2025 · 0 citations)
What looks settled
- Peptides-struct (MAE): best up to 2024 0.245, best in 2025–26 0.243 from 2 papers, less than 1% better.
- Pubmed (accuracy): best up to 2024 90.46, best in 2025–26 89.05 from 3 papers, at most 0.2% better, if at all.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2019 · Diffusion Improves Graph Learning (NeurIPS 2019 · 172 citations)
- 2020 · Dynamic Graph Message Passing Networks (CVPR 2020 · 146 citations)
- 2021 · E(n) Equivariant Graph Neural Networks (ICML 2021 · 102 citations)
- 2022 · Recipe for a General, Powerful, Scalable Graph Transformer (NeurIPS 2022 · 228 citations)
- 2023 · HOOD: Hierarchical Graphs for Generalized Modelling of Clothing Dynamics (CVPR 2023 · 82 citations)
- 2024 · Provably Powerful Graph Neural Networks for Directed Multigraphs (AAAI 2024 · 28 citations)
- 2025 · On Oversquashing in Graph Neural Networks Through the Lens of Dynamical Systems (AAAI 2025 · 3 citations)
- 2026 · Tab-PET: Graph-Based Positional Encodings for Tabular Transformers (AAAI 2026 · 2 citations)
Earliest papers here
- Diffusion Improves Graph Learning (NeurIPS 2019 · 172 citations)
- Invariant and Equivariant Graph Networks (ICLR 2019 · 98 citations)
- Hyper-Graph-Network Decoders for Block Codes (NeurIPS 2019 · 20 citations)
Benchmark records
The best number each year that beat every earlier one, among papers of this approach reporting their own method. Numbers are extracted by AI from each paper's text; check the paper before citing.
PROTEINS · accuracy (higher is better)
8 papers, 2019–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 75.19 | Invariant and Equivariant Graph Networks (ICLR 2019 · 98 citations) |
| 2021 | 76.3 | DropGNN: Random Dropouts Increase the Expressiveness of Graph Neural Networks (NeurIPS 2021 · 15 citations) |
| 2024 | 79.86 | Boosting Graph Pooling with Persistent Homology (NeurIPS 2024 · 6 citations) |
NCI1 · accuracy (higher is better)
6 papers, 2019–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 73.71 | Invariant and Equivariant Graph Networks (ICLR 2019 · 98 citations) |
| 2022 | 82.8 | SpeqNets: Sparsity-aware permutation-equivariant graph networks (ICML 2022 · 0 citations) |
| 2024 | 83.75 | Boosting Graph Pooling with Persistent Homology (NeurIPS 2024 · 6 citations) |
Peptides-struct · MAE (lower is better)
6 papers, 2022–2025 · small gains only
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 0.25 | Recipe for a General, Powerful, Scalable Graph Transformer (NeurIPS 2022 · 228 citations) |
| 2023 | 0.245 | A Generalization of ViT/MLP-Mixer to Graphs (ICML 2023 · 7 citations) |
| 2025 | 0.243 | Sketch-Augmented Features Improve Learning Long-Range Dependencies in Graph Neural Networks (NeurIPS 2025 · 0 citations) |
Cora · accuracy (higher is better)
5 papers, 2021–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2021 | 82.7 | Elastic Graph Neural Networks (ICML 2021 · 5 citations) |
| 2025 | 90.74 | GNNs Getting ComFy: Community and Feature Similarity Guided Rewiring (ICLR 2025 · 0 citations) |
Pubmed · accuracy (higher is better)
5 papers, 2021–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2021 | 79.4 | Elastic Graph Neural Networks (ICML 2021 · 5 citations) |
| 2023 | 90.46 | Tailoring Self-Attention for Graph via Rooted Subtrees (NeurIPS 2023 · 2 citations) |
Citeseer · accuracy (%) (higher is better)
4 papers, 2021–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2021 | 70.9 | Elastic Graph Neural Networks (ICML 2021 · 5 citations) |
| 2025 | 78.61 | Relieving the Over-Aggregating Effect in Graph Transformers (NeurIPS 2025 · 0 citations) |
What has been tried
74 papers here report 773 measured changes (a component added, removed or swapped, compared with a baseline): helped 69%, no clear effect 7%, hurt 17%, mixed 7%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add reverse message passing to GIN | Synthetic degree-out · minority-class F1 | 43.58 -> 99.08 | Helped | Provably Powerful Graph Neural Networks for Directed Multigraphs (AAAI 2024 · 28 citations) |
| replace uniform neighbor aggregation with separate transformations for fraud and benign neighbors | Yelp, Amazon, and T-Finance · AUC | 89.38 -> 93.15; 93.16 -> 96.33; 95.05 -> 95.92 | Helped | Partitioning Message Passing for Graph Fraud Detection (ICLR 2024 · 7 citations) |
| Add resampling | graph classification: NCI1, PROTEINS, ENZYMES, and IMDB-BINARY · accuracy | DiffPool NCI1: 80.82 ± 1.71 -> 83.75 ± 1.70; PROTEINS: 77.89 ± 4.07 -> 79.86 ± 3.12 | Helped | Boosting Graph Pooling with Persistent Homology (NeurIPS 2024 · 6 citations) |
| add the flow encode module | NAS-Bench-101 · Kendall's Tau (×100) | 42.5 (1.6) -> 86.1 (0.8) | Helped | FlowerFormer: Empowering Neural Architecture Encoding using a Flow-aware Graph Transformer (CVPR 2024 · 5 citations) |
| Replace α-entmax sampling with Gumbel Top-k | Homophilic and heterophilic datasets · accuracy | Hurt | From Latent Graph to Latent Topology Inference: Differentiable Cell Complex Module (ICLR 2024 · 4 citations) | |
| Add spatial antisymmetry (use beta instead of beta=0) | Peptides-func · average precision | 61.95 (±00.67) -> 63.13 (±0.46) | Helped | On Oversquashing in Graph Neural Networks Through the Lens of Dynamical Systems (AAAI 2025 · 3 citations) |
| Use a sum-aggregation action network instead of a mean-aggregation action network | RootNeighbors · MAE | 0.339 -> 0.079 | Helped | Cooperative Graph Neural Networks (ICML 2024 · 3 citations) |
| Add LapPE to GCN+Degree | Topology Far-OOD · accuracy | 33.67 ± 3.56 -> 55.00 ± 3.08 | Helped | The Underappreciated Power of Vision Models for Graph Structural Understanding (NeurIPS 2025 · 1 citation) |
| Add k-hop aggregation to SSM dynamics | GraphProp Diameter, SSSP, and Eccentricity · log10(MSE) | Diameter: -2.431 -> -3.075; SSSP: -2.821 -> -3.604; Eccentricity: -2.245 -> -4.265 | Helped | On Vanishing Gradients, Over-Smoothing, and Over-Squashing in GNNs: Bridging Recurrent and Graph Learning (NeurIPS 2025 · 1 citation) |
| add identity filtration (VC-I) | ZINC · MAE | 0.1878 ±\textcolor{gray}{0.012} -> 0.1432 ±\textcolor{gray}{0.023} | Helped | Positional Encoding meets Persistent Homology on Graphs (ICML 2025 · 0 citations) |
| use joint density estimation over all layer representations rather than estimating density from one layer at a time | Amazon Ratings and Roman Empire, LoC · AUC-ROC | Helped | Uncertainty Estimation for Heterophilic Graphs Through the Lens of Information Theory (ICML 2025 · 0 citations) | |
| Increase the training label ratio | CSBM and Cora · accuracy | null | Helped | A Signed Graph Approach to Understanding and Mitigating Oversmoothing (NeurIPS 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 28 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- Cora · 8 of 28 papers
- amazon-ratings · 5 of 28 papers
- PubMed · 5 of 28 papers
- PROTEINS · 5 of 28 papers
- Citeseer · 5 of 28 papers
- Peptides-struct · 4 of 28 papers
- ZINC · 4 of 28 papers
- roman-empire · 4 of 28 papers
Baselines
- GCN · 11 of 28 papers
- GIN · 8 of 28 papers
- GAT · 7 of 28 papers
- GraphSAGE · 6 of 28 papers
- GPS · 4 of 28 papers
- GCNII · 3 of 28 papers
- SignNet · 3 of 28 papers
- SPE · 3 of 28 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 83 papers; from OpenAlex and the papers' affiliations).
- University of Oxford · 5 papers
- Peking University · 3 papers
- Google · 3 papers
- Aalto University · 3 papers
- Technical University Munich · 3 papers
- The Chinese University of Hong Kong · 3 papers
- IBM · 2 papers
- University of Pisa · 2 papers
- University of Cambridge · 2 papers
- University of Science and Technology of China · 2 papers
Use it
Download BibTeX of 39 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Graph rewiring and scalable message passing (72 papers)
- Structural GNN variants (65 papers)
- Equivariant networks for molecules (43 papers)
Elsewhere in the topic
- Graph structure learning (201 papers)
- GNN generalization theory (187 papers)
- Simple and scalable GNNs (131 papers)
- Robustness, privacy, and fairness of GNNs (94 papers)
- Theory of graph convolutional networks (83 papers)
- Deep graph convolution and simplification (82 papers)
- Physics-inspired GNNs (81 papers)
- Weisfeiler-Leman expressivity hierarchies (66 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).